# Stress Transcript Variants

The precise stress comes from the LIEPA synthesis system, not from the included PDF.

The four voice recordings were meticulously marked for stress during development of the original LIEPA voices. Those unpublished annotations were used to derive the synthesizer's statistical stress-selection model. The exact recorded prompts were later processed with the modified open-source LIEPA text processor, published as [`phonology_engine`](https://github.com/aleksas/phonology_engine). Its output therefore reconstructs the intended stress associated with these recordings.

This is an engine-based reconstruction, not a direct export of the original manual annotations, which were never published.

## Variants

- **Variant A** (`text_stressed_ascii_a`, `text_stressed_unicode_a`): engine-reconstructed stress. The Unicode form is recommended and is also exposed as `phonetic_transcription`.
- **Numeric** (`text_stressed_numbers`): the same engine-derived stress in the LIEPA synthesizer's internal encoding.
- **Variant B** (`text_stressed_ascii_b`, `text_stressed_unicode_b`): stress transcribed from `atrinkti_sak_2a.pdf`. The PDF is useful for spelling and capitalization but its stress notation is less precise.

The generated speaker JSON files align these sources through `id` and `ref_id`. Edvardas's source transcript lacks one PDF sentence (reference ID 2435).

## Stress Marks

| Stress | ASCII | Unicode combining |
|---|---|---|
| Grave | `` ` `` | `\u0300` |
| Acute | `^` | `\u0301` |
| Tilde | `~` | `\u0303` |